WhyTrend Bitcoin Anomaly Analysis: Smaller Model Outperforms Larger One

WhyTrend, an open-source pipeline for anomaly detection and explanation, was run on daily Bitcoin close prices (BTC/USDT) from January 1, 2022 to August 7, 2026, totaling 1680 data points. The detector found 8 changepoints (regime changes), which were then explained using two different local LLMs: Qwen 32B and qwen2.5-coder:14b. The larger 32B model frequently timed out (6 out of 8 events), causing fallback to heuristic explanations based on top BM25-ranked articles, leading to nonsensical reasons like 'Stray Cats and Bitcoin' or 'Bitcoin Genetic Code Paper'. The smaller 14B model completed within timeout for 6 out of 8 events, providing more credible explanations with confidence scores, such as attributing the 2024-02-25 changepoint to negative sentiment and regulatory concerns, and the 2025-11-11 changepoint to governments seizing Bitcoin and accusations of a US-backed $13B hack. The article notes that while 14B is on average more reliable, it is not perfect, as it also had fallback cases. The author recommends using smaller models for such tasks and discusses the importance of distinguishing fallback outputs from genuine model reasoning.